Why does GPT-4-vision-preview always report it cannot see images?

Below is some guidance on issues that may be causing your reported “sorry, I can’t see images” response.

---

1. Use the Correct Model Identifier

The OpenAI documentation clarifies that for image understanding you need to use the specific vision variants. The snippet provided in the docs uses models such as gpt-4o-mini (or others like gpt-4-turbo). Using the placeholder or experimental identifier like gpt-4-vision-preview may result in a model that isn’t configured to process images. Make sure you use one of the documented models for vision tasks. For example:

PYTHONCopied
model="gpt-4o-mini"

---

2. Verify Image Accessibility

Ensure that the image URL you are passing is publicly accessible to OpenAI’s servers. If the image is behind any sort of access control or not served over HTTPS, the model won’t be able to fetch and process the image. You might also consider converting the image to base64 and sending that instead, as some integrations require that.

---

3. Match the API Endpoint and Request Format

The official examples for image analysis combine text and image inputs using a chat completion endpoint. That ensures the image is correctly interpreted. For example, the intended request structure is similar to:

PYTHONCopied
import openai

response = openai.ChatCompletion.create(
    model="gpt-4o-mini",
    messages=[
       {
           "role": "user",
           "content": [
             {"type": "text", "text": "What’s in this image?"},
             {
               "type": "image_url",
               "image_url": {"url": "https://your-accessible-image-url.com/image.png"}
             }
          ]
       }
    ],
    max_tokens=300,
)

Using the correct endpoint and message structure is necessary for proper interpretation of the image.

---

4. Confirm API Access and Permissions

Your API key and account need to have access to the vision-enabled models. Make sure the account is enabled for these capabilities and you’re using the appropriate endpoints as described in the OpenAI docs. The “vision” feature is more sensitive and sometimes gated.

---

Summary

• Model Selection: Switch from gpt-4-vision-preview to one of the designated image understanding models (like gpt-4o-mini).
• Image Accessibility: Confirm that the image URL is publicly accessible or try using base64 encoding.
• Request Format: Use the documented request structure for combining text with image input.
• Permissions: Verify that your API key/account has access to the vision model variants.

Adjust these factors in your code, and you should be able to get a more descriptive evaluation of your DALL·E 3 generated image.